AI at Work: Measure Time Saved, Not the Number of Uses

What the news says, and what it does not allow us to conclude

The publication of the Banque de France describes a wide diffusion of generative AI: more than two thirds of companies with at least twenty employees declare using it at the beginning of 2026. It emphasizes, however, that the economic effects observed remain limited at this stage. This scope does not allow this figure to be presented as the adoption rate of all French small businesss.

Our practical reading is as follows: counting the people equipped is not enough to decide on an investment. The number of conversations with an assistant reveals neither the quality of the deliverables nor the time spent resuming them. These indicators can describe an activity; they do not directly measure a benefit.

For a manager, the useful question therefore becomes: “On what specific task do we obtain an acceptable result with less total work?” » This wording invites examination of the organization, available data and verification. It also allows us to see that a use is pleasant without being profitable yet.

Choose a task whose completion can be verified

Start with repetitive work, frequent enough to observe several cases and whose result can be controlled. Preparing a draft response to a simple request may be appropriate. Entrusting a binding business decision unchecked constitutes another level of responsibility and does not serve as a starting point for this exercise.

Define what “finished” means. Is a generated text just a draft? Should it be checked in the catalog, corrected and then integrated into your software? The timer should stop when the document reaches the expected level, not when the first sentences appear on the screen.

• Describe the task in one sentence and limit its scope.
• Choose cases comparable in difficulty and length.
• Write the information that must be accurate.
• Identify the person capable of validating the result.
• Maintain a way to process the file without the wizard.

If the process itself remains unclear, start by clarifying it. Our guide to implement AI in a small business or small business details this framing. Here, the objective is more precise: to construct a measurement that your team can repeat without interpreting each file differently.

A fictitious example: control time changes the balance sheet

Let's imagine twenty simple answers to prepare, with eight minutes per answer in usual operation. The prize represents one hundred and sixty minutes. With a wizard, assume three minutes of preparation and generation per response, two minutes of review, and thirty minutes spent on initial setup.

The winnings for this first prize would be thirty minutes. This calculation is entirely fictitious: it illustrates a method, not a customer result or an average performance. In your actual measurement, the durations and distribution of control may be very different.

On a following identical batch, without new configuration, the hypothesis would lead to one hundred minutes of work. But don't automatically delete thirty minutes from all your calculations: instructions may evolve, models change and a new person may need training. Note these time expenditures as they arise.

If the check ultimately takes six minutes per response, the same first batch reaches two hundred and ten minutes. The tool then speeds up the initial writing while lengthening the total work. This example shows why a quick demonstration is no substitute for end-to-end measurement.

Measuring quality alongside speed

A faster result can be costly if it contains an incorrect reference or a commitment that the company cannot keep. Prepare a short grid: accuracy of the facts, compliance with the request, absence of sensitive information and possibility of using the document after validation. A missing essential criterion must remain visible.

You can classify each output into three states: usable after checking, to be resumed, to be redone. Note the pattern of repeats. A recurring correction on deadlines may reveal an incomplete instruction or an obsolete reference document. This observation is more useful than a general rating of satisfaction.

To avoid a biased comparison, do not only give the easy cases to the assistant. Keep a mix representative of your activity, within the limits of the authorized perimeter. If an unusual file requires human intervention, count this intervention in the balance rather than silently removing the file from the trial.

Our method for check an AI response helps organize this proofreading. The important thing is to make the verification repeatable: two people must understand the same criteria and know when to request additional validation.

Protect information during testing

The France Num sheet on precautions for using AI, updated on September 9, 2026, recalls the interest of a framework of use, verification of results and protection of confidential information. These are conditions of the test, not tasks to be postponed until after its generalization.

Prepare fictional or properly anonymized examples for a first try when possible. Removing a name is not always enough: an address, file reference or combination of details can still identify a customer. Check what is actually transmitted and the conditions of the service used.

Also write the uses excluded from the experiment. A team needs to know if it can prepare generic sales text, analyze an internal document, or import a customer file. A concrete rule, accompanied by a contact person in case of doubt, is easier to apply than a general invitation to be careful.

Decide: continue, modify or stop

At the end of the chosen period, gather the total time, rework, tool costs and difficulties observed. Compare work of equivalent quality. Separate upfront costs from recurring expenses to understand what could change if usage continues.

Freed up time is not automatically a saving in your bank account. It can be used to respond more quickly, to absorb a peak in activity or to carry out work that had previously been postponed. Indicate the intended use of this time before converting minutes into euros and announcing profitability.

Continue when the benefit is observable and control remains under control. Modify the test if the fixes focus on a specific problem. Stop it if the quality or processing constraints make the use unsuitable. A simple automation for small business, based on rules, can also meet the need without text generation.

Frequently asked questions about the benefits of AI in business

How many files should be tested?

There is no universal number suitable for all activities. Choose enough comparable cases to observe the usual variety of work. The twenty answers in our example are used solely for educational calculations and do not constitute a statistical rule.

Should we count the time spent writing the instructions?

Yes. Include initial preparation, then modifications and training as they occur. You can present upfront costs and recurring times separately, while still maintaining a complete total.

Can we measure the gain only with the number of texts produced?

This number provides information on the volume. Add the control time, retakes and actual use of texts to understand their usefulness. A document produced and then abandoned should not be counted as a final validated deliverable.

What to do if the team likes the tool without saving time?

Examine what she appreciates: help getting started, comfort, better structure or less difficulty on a task. This benefit may have value, but name it correctly and compare it to the cost. It does not prove an increase in productivity.